unitysvc-data 0.2.6__tar.gz → 0.2.8__tar.gz

{unitysvc_data-0.2.6/src/unitysvc_data.egg-info → unitysvc_data-0.2.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: unitysvc-data
-Version: 0.2.6
+Version: 0.2.8
 Summary: Standard examples and presets for UnitySVC data packages
 Author-email: Bo Peng <bo.peng@unitysvc.com>
 Maintainer-email: Bo Peng <bo.peng@unitysvc.com>

{unitysvc_data-0.2.6 → unitysvc_data-0.2.8}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "unitysvc-data"
-version = "0.2.6"
+version = "0.2.8"
 description = "Standard examples and presets for UnitySVC data packages"
 readme = "README.md"
 authors = [{ name = "Bo Peng", email = "bo.peng@unitysvc.com" }]

{unitysvc_data-0.2.6 → unitysvc_data-0.2.8}/src/unitysvc_data/_manifest.json

@@ -99,8 +99,7 @@
     "llm_connectivity_tts": "llm_connectivity_tts_v1",
     "llm_connectivity_tts_dashscope": "llm_connectivity_tts_dashscope_v1",
     "llm_description": "llm_description_v2",
-    "llm_request_template": "llm_request_template_v1",
-    "llm_request_template_anthropic": "llm_request_template_anthropic_v1",
+    "llm_request_template": "llm_request_template_v2",
     "msg_request_template": "msg_request_template_v1",
     "msg_to_apprise_code_example_py": "msg_to_apprise_code_example_py_v1",
     "msg_to_apprise_code_example_sh": "msg_to_apprise_code_example_sh_v1",
@@ -3765,31 +3764,29 @@
       "source_readme": "llm/description/README.md",
       "version": 2
     },
-    "llm_request_template_anthropic_v1": {
+    "llm_request_template_v1": {
       "applies_to": {
-        "capability": "chat",
-        "upstream": "anthropic"
+        "capability": "chat"
       },
       "category": "request_template",
-      "description": "Minimal Anthropic Messages API request body",
-      "example_file": "llm/request-template-anthropic/request-template-anthropic-v1.json",
+      "description": "Minimal chat request body for each request format",
+      "example_file": "llm/request-template/request-template-v1.json",
       "is_active": true,
       "is_public": true,
       "meta": {},
       "mime_type": "json",
       "parameters": {},
-      "preset_name": "llm_request_template_anthropic",
-      "source_readme": "llm/request-template-anthropic/README.md",
+      "preset_name": "llm_request_template",
+      "source_readme": "llm/request-template/README.md",
       "version": 1
     },
-    "llm_request_template_v1": {
+    "llm_request_template_v2": {
       "applies_to": {
-        "capability": "chat",
-        "upstream": "openai"
+        "capability": "chat"
       },
       "category": "request_template",
-      "description": "Minimal OpenAI-compatible chat completion request body",
-      "example_file": "llm/request-template/request-template-v1.json",
+      "description": "Minimal chat request body for each request format",
+      "example_file": "llm/request-template/request-template-v2.json",
       "is_active": true,
       "is_public": true,
       "meta": {},
@@ -3797,7 +3794,7 @@
       "parameters": {},
       "preset_name": "llm_request_template",
       "source_readme": "llm/request-template/README.md",
-      "version": 1
+      "version": 2
     },
     "msg_request_template_v1": {
       "applies_to": {},

{unitysvc_data-0.2.6 → unitysvc_data-0.2.8}/src/unitysvc_data/examples/llm/dashscope-omni-tts-javascript/code-example-v1.js.j2

@@ -1,4 +1,6 @@
 import { writeFile } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
 
 async function main() {
   // Audio output is streaming-only: without the SSE header the call still
@@ -38,7 +40,9 @@ async function main() {
     }
   }
 
-  const outputFile = process.env.OUTPUT_FILE ?? "speech.wav";
+  // Into a temp dir by default: the test runner executes this from a
+  // working directory it cannot write to.
+  const outputFile = process.env.OUTPUT_FILE ?? join(tmpdir(), "speech.wav");
   await writeFile(outputFile, Buffer.from(chunks.join(""), "base64"));
   console.log(`Saved ${outputFile} (${chunks.length} audio frame(s))`);
 {%- if not customer_display %}

{unitysvc_data-0.2.6 → unitysvc_data-0.2.8}/src/unitysvc_data/examples/llm/dashscope-omni-tts-requests/code-example-v1.py.j2

@@ -1,6 +1,7 @@
 import base64
 import json
 import os
+import tempfile
 
 import requests
 
@@ -35,7 +36,12 @@ for line in response.iter_lines(decode_unicode=True):
             if isinstance(audio, dict) and audio.get("data"):
                 chunks.append(audio["data"])
 
-output_file = os.environ.get("OUTPUT_FILE", "speech.wav")
+# Into a temp dir by default: the test runner executes this from a
+# working directory it cannot write to, and a PermissionError there
+# fails a script whose request, stream and decode all succeeded.
+output_file = os.environ.get("OUTPUT_FILE") or os.path.join(
+    tempfile.gettempdir(), "speech.wav"
+)
 with open(output_file, "wb") as f:
     f.write(base64.b64decode("".join(chunks)))
 print(f"Saved {output_file} ({len(chunks)} audio frame(s))")

{unitysvc_data-0.2.6 → unitysvc_data-0.2.8}/src/unitysvc_data/examples/llm/dashscope-omni-tts-shell/code-example-v1.sh.j2

@@ -2,7 +2,9 @@
 set -e -o pipefail
 
 : "${UNITYSVC_API_KEY:?UNITYSVC_API_KEY is not set}"
-output_file="${OUTPUT_FILE:-speech.wav}"
+# Into a temp dir by default: the test runner executes this from a
+# working directory it cannot write to.
+output_file="${OUTPUT_FILE:-${TMPDIR:-/tmp}/speech.wav}"
 
 # Audio output is streaming-only, hence the SSE header: without it the call
 # succeeds, bills audio tokens, and returns no audio.

unitysvc_data-0.2.8/src/unitysvc_data/examples/llm/request-template/README.md

@@ -0,0 +1,85 @@
++++
+preset_name = "llm_request_template"
+category = "request_template"
+mime_type = "json"
+file = "request-template.json"
+description = "Minimal chat request body for each request format"
+is_active = true
+is_public = true
+applies_to = { capability = "chat" }
++++
+
+# llm / request-template — minimal chat payload, per request format
+
+The default request body the Test Request playground offers for a chat
+service. From v2 it carries one body per request format, keyed by the
+gateway's format names (`input_formats` / `request_formats`), so every
+chat service ships the same document whatever formats it accepts. The
+playground uses `template[format]` for the format the customer picked
+(unitysvc/unitysvc#2508); entries for formats a service doesn't accept are
+ignored.
+
+## Body (v2)
+
+| Key                | Wire shape                                     |
+|--------------------|------------------------------------------------|
+| `openai`           | OpenAI Chat Completions                        |
+| `anthropic`        | Anthropic Messages (`system` is top-level)     |
+| `cohere`           | Cohere v2 Chat (native `/v2/chat`)             |
+| `dashscope`        | DashScope native (`input.messages` + `parameters`) |
+| `bedrock_converse` | Bedrock Converse (content blocks, `inferenceConfig`) |
+
+```json
+{
+  "openai":    { "max_tokens": 100, "messages": [ system, user ] },
+  "anthropic": { "max_tokens": 100, "system": "...", "messages": [ user ] },
+  "cohere":    { "max_tokens": 100, "messages": [ system, user ] },
+  "dashscope": { "input": { "messages": [ system, user ] },
+                 "parameters": { "max_tokens": 100, "result_format": "message" } },
+  "bedrock_converse": { "system": [ { "text": "..." } ],
+                        "messages": [ { "role": "user", "content": [ { "text": "..." } ] } ],
+                        "inferenceConfig": { "maxTokens": 100 } }
+}
+```
+
+Every entry says the same thing: system prompt "You are a helpful
+assistant.", user prompt "Say hello in one sentence.", at most 100 tokens.
+
+## What's intentionally missing
+
+- **No `model` field.** The gateway's routing config or the listing's
+  `upstream_access_config` selects the upstream model; the playground merges
+  the service's routing key into the body it sends. Bedrock Converse takes the
+  model id in the path instead.
+- **No `stream` field.** Non-streaming responses are simpler to
+  validate in tests; streaming variants belong in separate presets.
+- **No `bedrock_invoke` entry.** InvokeModel's body is the model family's own
+  (Anthropic-on-Bedrock, Llama, Titan, …), so there is no one body to offer.
+  The playground falls back to a blank body for it.
+
+## Conventions
+
+- `max_tokens` is kept small (≤ 100) so the request completes fast
+  against any upstream regardless of per-token latency.
+- A new request format gets a new key here, in a new version.
+
+## Versions
+
+### v2 — one body per request format
+
+- Keyed by format: `openai`, `anthropic`, `cohere`, `dashscope`,
+  `bedrock_converse`. The `openai` entry is v1's body.
+- `applies_to` no longer requires an OpenAI upstream, so every chat service gets
+  this template whatever its dialect — which is what makes one keyed body the
+  right shape: a service accepting both OpenAI and Anthropic needs ONE default
+  body, and the playground indexes it by the format the customer picked.
+  `llm_request_template_anthropic` is removed rather than kept alongside; per-format
+  templates would give such a service two `Default request body` documents and no
+  way to say which is the default.
+- Needs a playground that picks a format's entry (unitysvc/unitysvc#2508).
+  An older one shows the whole map as the body.
+
+### v1 — initial release
+
+- Two messages (system + user), `max_tokens = 100`, no model field,
+  non-streaming.

unitysvc_data-0.2.8/src/unitysvc_data/examples/llm/request-template/request-template-v2.json

@@ -0,0 +1,76 @@
+{
+  "openai": {
+    "max_tokens": 100,
+    "messages": [
+      {
+        "role": "system",
+        "content": "You are a helpful assistant."
+      },
+      {
+        "role": "user",
+        "content": "Say hello in one sentence."
+      }
+    ]
+  },
+  "anthropic": {
+    "max_tokens": 100,
+    "system": "You are a helpful assistant.",
+    "messages": [
+      {
+        "role": "user",
+        "content": "Say hello in one sentence."
+      }
+    ]
+  },
+  "cohere": {
+    "max_tokens": 100,
+    "messages": [
+      {
+        "role": "system",
+        "content": "You are a helpful assistant."
+      },
+      {
+        "role": "user",
+        "content": "Say hello in one sentence."
+      }
+    ]
+  },
+  "dashscope": {
+    "input": {
+      "messages": [
+        {
+          "role": "system",
+          "content": "You are a helpful assistant."
+        },
+        {
+          "role": "user",
+          "content": "Say hello in one sentence."
+        }
+      ]
+    },
+    "parameters": {
+      "max_tokens": 100,
+      "result_format": "message"
+    }
+  },
+  "bedrock_converse": {
+    "system": [
+      {
+        "text": "You are a helpful assistant."
+      }
+    ],
+    "messages": [
+      {
+        "role": "user",
+        "content": [
+          {
+            "text": "Say hello in one sentence."
+          }
+        ]
+      }
+    ],
+    "inferenceConfig": {
+      "maxTokens": 100
+    }
+  }
+}

{unitysvc_data-0.2.6 → unitysvc_data-0.2.8}/src/unitysvc_data/presets.py

@@ -32,7 +32,6 @@ from __future__ import annotations
 
 import atexit
 import hashlib
-import itertools
 import json
 import os
 import tempfile
@@ -634,35 +633,44 @@ def llm_example_collection(source: Any) -> dict[str, Any]:
     # declares that it applies here. The probe lands on the primary group
     # because it must resolve to exactly one channel/interface; everything
     # else is scoped to the group that contributed it.
-    for capability, group in itertools.product(capabilities, groups):
-        features = {"streaming"}
-        if group.get("tools"):
-            features.add("tools")
-        # `vision` is implied by the capability that needs it, so a listing
-        # declaring `image-text-to-text` does not also have to set the flag.
-        # The flag is still honoured on its own for the callers that set it.
-        if group.get("vision") or capability in _VISION_CAPABILITIES:
-            features.add("vision")
-        for title, preset_name in _select(
-            capability=capability,
-            dialects=set(group["formats"]),
-            # Which dialect the UPSTREAM speaks is a property of the channel a
-            # group maps to (its ``upstream_format``), not of the service: one
-            # listing can front an OpenAI-compatible channel and a native one
-            # at the same time (QwenCloud's compat layer plus DashScope, where
-            # only the native surface serves TTS/ASR). The collection-level
-            # value stays the default, so every existing repo is unaffected.
-            upstream=group.get("upstream_dialect") or upstream,
-            features=features,
-        ):
-            scope = _primary_group(groups) if title == "Connectivity test" else group
-            docs[title] = _scoped(
-                preset_name,
-                scope,
-                sleep,
-                {**broadcast, **(group.get("params") or {})},
-                min_expected_metrics=min_expected_metrics,
-            )
+    # Per GROUP, then per capability, because a group may serve only a SUBSET.
+    # `capabilities` fanned across every group, so each advertised the service's
+    # whole list whether or not its endpoint served them: QwenCloud's
+    # compatible-mode endpoint serves chat and embeddings but answers
+    # /v1/audio/speech and /v1/audio/transcriptions with 404, and its compat group
+    # still pulled in the stock OpenAI audio examples and failed them.
+    #
+    # Absent, a group serves all of them, so no existing repo changes.
+    for group in groups:
+        for capability in (group.get("capabilities") or capabilities):
+            features = {"streaming"}
+            if group.get("tools"):
+                features.add("tools")
+            # `vision` is implied by the capability that needs it, so a listing
+            # declaring `image-text-to-text` does not also have to set the flag.
+            # The flag is still honoured on its own for the callers that set it.
+            if group.get("vision") or capability in _VISION_CAPABILITIES:
+                features.add("vision")
+            for title, preset_name in _select(
+                capability=capability,
+                dialects=set(group["formats"]),
+                # Which dialect the UPSTREAM speaks is a property of the channel a
+                # group maps to (its ``upstream_format``), not of the service: one
+                # listing can front an OpenAI-compatible channel and a native one
+                # at the same time (QwenCloud's compat layer plus DashScope, where
+                # only the native surface serves TTS/ASR). The collection-level
+                # value stays the default, so every existing repo is unaffected.
+                upstream=group.get("upstream_dialect") or upstream,
+                features=features,
+            ):
+                scope = _primary_group(groups) if title == "Connectivity test" else group
+                docs[title] = _scoped(
+                    preset_name,
+                    scope,
+                    sleep,
+                    {**broadcast, **(group.get("params") or {})},
+                    min_expected_metrics=min_expected_metrics,
+                )
     return docs
 
 

{unitysvc_data-0.2.6 → unitysvc_data-0.2.8/src/unitysvc_data.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: unitysvc-data
-Version: 0.2.6
+Version: 0.2.8
 Summary: Standard examples and presets for UnitySVC data packages
 Author-email: Bo Peng <bo.peng@unitysvc.com>
 Maintainer-email: Bo Peng <bo.peng@unitysvc.com>

{unitysvc_data-0.2.6 → unitysvc_data-0.2.8}/src/unitysvc_data.egg-info/SOURCES.txt

@@ -242,8 +242,7 @@ src/unitysvc_data/examples/llm/description/description-v1.md
 src/unitysvc_data/examples/llm/description/description-v2.md
 src/unitysvc_data/examples/llm/request-template/README.md
 src/unitysvc_data/examples/llm/request-template/request-template-v1.json
-src/unitysvc_data/examples/llm/request-template-anthropic/README.md
-src/unitysvc_data/examples/llm/request-template-anthropic/request-template-anthropic-v1.json
+src/unitysvc_data/examples/llm/request-template/request-template-v2.json
 src/unitysvc_data/examples/msg-to-apprise/code-example-py/README.md
 src/unitysvc_data/examples/msg-to-apprise/code-example-py/code-example-v1.py.j2
 src/unitysvc_data/examples/msg-to-apprise/code-example-sh/README.md

{unitysvc_data-0.2.6 → unitysvc_data-0.2.8}/tests/test_example_collection.py

@@ -10,6 +10,7 @@ carry its preset name, so the path is not a usable key.
 
 from __future__ import annotations
 
+import json
 import re
 from pathlib import Path
 
@@ -243,13 +244,61 @@ def test_chat_collection_carries_the_description_and_request_template():
     assert presets_in(docs) >= {"llm_description", "llm_request_template"}
 
 
-def test_anthropic_upstream_gets_the_anthropic_request_template():
-    docs = llm_example_collection(
-        {"capabilities": ["chat"], "formats": ["anthropic"], "upstream_dialect": "anthropic"}
+def test_every_chat_upstream_gets_the_one_format_keyed_request_template():
+    """v2 keys a body per request format, so an Anthropic upstream no longer
+    needs its own template — and must not get both, which would collide on
+    the title "Default request body"."""
+    for upstream in ("openai", "anthropic"):
+        docs = llm_example_collection(
+            {"capabilities": ["chat"], "formats": ["anthropic"], "upstream_dialect": upstream}
+        )
+        templates = [d for d in docs.values() if d["category"] == "request_template"]
+        assert len(templates) == 1
+        assert "llm_request_template" in presets_in(docs)
+
+
+def test_the_request_template_carries_a_body_for_each_request_format():
+    body = json.loads(
+        Path(doc_preset("llm_request_template")["file_path"]).read_text()
     )
+    assert set(body) == {"openai", "anthropic", "cohere", "dashscope", "bedrock_converse"}
+    # The gateway recognises each body as its own format by shape
+    # (apisix-gateways request_meta.classify); keep the markers it keys on.
+    assert "system" in body["anthropic"] and "max_tokens" in body["anthropic"]
+    assert isinstance(body["dashscope"]["input"], dict)
+    assert body["bedrock_converse"]["messages"][0]["content"] == [
+        {"text": "Say hello in one sentence."}
+    ]
+    # No entry names a model: the routing key is merged in by the caller.
+    assert all("model" not in entry for entry in body.values())
 
-    assert "llm_request_template_anthropic" in presets_in(docs)
-    assert "llm_request_template" not in presets_in(docs)
+
+def test_the_pinned_v1_template_is_unchanged():
+    v1 = json.loads(
+        Path(doc_preset("llm_request_template_v1")["file_path"]).read_text()
+    )
+    assert set(v1) == {"max_tokens", "messages"}
+
+
+def test_exactly_one_request_template_preset_serves_chat():
+    """One keyed body, not one preset per format. A service accepting both
+    OpenAI and Anthropic needs ONE default request body — the playground indexes
+    it by the format the customer picked — and per-format presets would hand it
+    two `Default request body` documents with no way to say which is the default.
+
+    `llm_request_template_anthropic` was removed rather than kept unselected, so
+    this asserts nothing has reintroduced a second one."""
+    from unitysvc_data import PRESETS
+
+    llm_templates = sorted(
+        n for n in PRESETS if "request_template" in n and n.startswith("llm_")
+    )
+    # The family plus its version aliases, and nothing else.
+    assert llm_templates == [
+        "llm_request_template",
+        "llm_request_template_v1",
+        "llm_request_template_v2",
+    ], llm_templates
 
 
 def test_sleep_is_applied_to_every_executable_document():
@@ -687,8 +736,7 @@ def test_the_request_template_is_selected_like_everything_else():
     chat — rather than the collection special-casing chat to add it."""
     from unitysvc_data import applies_to
 
-    assert applies_to("llm_request_template")["capability"] == "chat"
-    assert applies_to("llm_request_template_anthropic")["upstream"] == "anthropic"
+    assert applies_to("llm_request_template") == {"capability": "chat"}
 
 
 def test_the_how_to_doc_declares_nothing_because_it_is_universal():
@@ -1062,15 +1110,22 @@ def test_every_capability_example_is_pinned_to_a_wire_shape():
     `dialect` pins what the CALLER writes, `upstream` what the upstream speaks,
     and a DashScope channel differs on both. `llm_description` is the deliberate
     exception — it constrains nothing because it really does apply to everything.
+
+    `llm_request_template` is the other one: from v2 it carries a body for EVERY
+    request format, keyed by format name, and the playground picks the entry for
+    the format the caller writes. It is pinned to a wire shape per entry rather
+    than per document.
     """
     from unitysvc_data import MANIFEST
     from unitysvc_data.presets import applies_to
 
+    keyed_by_format = {"llm_request_template"}
     names = {r.get("preset_name", k) for k, r in MANIFEST["presets"].items()
              if k.startswith("llm_")}
     gaps = sorted(n for n in names
                   if (a := applies_to(n)).get("capability")
-                  and not a.get("dialect") and not a.get("upstream"))
+                  and not a.get("dialect") and not a.get("upstream")
+                  and n not in keyed_by_format)
     assert gaps == [], f"these name a capability but no wire shape: {gaps}"
 
 
@@ -1183,3 +1238,39 @@ def test_omni_audio_floors_are_tokens_not_characters():
         floor = (doc.get("meta") or {}).get("min_expected_metrics")
         if floor is not None:
             assert "characters" not in floor, f"{title} still demands characters"
+
+
+def test_a_group_may_serve_a_subset_of_the_service_s_capabilities():
+    """`capabilities` fans across every group, so a group advertises the service's
+    whole capability list whether or not its endpoint serves them. QwenCloud's
+    compatible-mode endpoint serves chat and embeddings but answers
+    `/v1/audio/speech` and `/v1/audio/transcriptions` with 404 — yet an omni
+    service's compat group pulled in the stock OpenAI audio examples and failed
+    them, while its native group handled audio correctly."""
+    docs = llm_example_collection({
+        "capabilities": ["chat", "text-to-speech"],
+        "formats": [
+            # The compat endpoint serves chat only; say so.
+            {"formats": ["openai"], "capabilities": ["chat"],
+             "interface": "canonical", "primary": True},
+            {"formats": ["dashscope_multimodal"], "interface": "dashscope",
+             "upstream_dialect": "dashscope"},
+        ],
+    })
+    names = examples_in(docs)
+    # The stock /v1/audio/speech example must NOT be selected...
+    assert "llm_code_example_tts_requests" not in names
+    # ...while the native one is, and chat still reaches the compat endpoint.
+    assert "llm_code_example_omni_tts_dashscope_requests" in names
+    assert "llm_code_example_requests" in names
+
+
+def test_a_group_without_capabilities_still_serves_all_of_them():
+    """Opt-in per group: every existing repo declares none and is unaffected."""
+    docs = llm_example_collection({
+        "capabilities": ["chat", "text-to-speech"],
+        "formats": [{"formats": ["openai"], "interface": "canonical"}],
+    })
+    names = examples_in(docs)
+    assert "llm_code_example_requests" in names
+    assert "llm_code_example_tts_requests" in names

unitysvc_data-0.2.6/src/unitysvc_data/examples/llm/request-template/README.md

@@ -1,51 +0,0 @@
-+++
-preset_name = "llm_request_template"
-category = "request_template"
-mime_type = "json"
-file = "request-template.json"
-description = "Minimal OpenAI-compatible chat completion request body"
-is_active = true
-is_public = true
-applies_to = { capability = "chat", upstream = "openai" }
-+++
-
-# llm / request-template — minimal chat completion payload
-
-Canonical JSON request body for OpenAI-compatible chat completion
-services routed through the UnitySVC LLM gateway. Suitable as a test
-payload for seller-side validation, and as `request_template`
-metadata attached to a listing.
-
-## Body
-
-```json
-{
-  "max_tokens": 100,
-  "messages": [
-    { "role": "system", "content": "You are a helpful assistant." },
-    { "role": "user",   "content": "Say hello in one sentence." }
-  ]
-}
-```
-
-## What's intentionally missing
-
-- **No `model` field.** The gateway's routing config or the listing's
-  `upstream_access_config` selects the upstream model; hard-coding
-  one here would tie the template to a specific service.
-- **No `stream` field.** Non-streaming responses are simpler to
-  validate in tests; streaming variants belong in separate presets.
-
-## Conventions
-
-- `max_tokens` is kept small (≤ 100) so the request completes fast
-  against any upstream regardless of per-token latency.
-- Response format is standard OpenAI `chat.completions`. Tests should
-  assert `choices[0].message.content` exists and is non-empty.
-
-## Versions
-
-### v1 — initial release
-
-- Two messages (system + user), `max_tokens = 100`, no model field,
-  non-streaming.

unitysvc_data-0.2.6/src/unitysvc_data/examples/llm/request-template-anthropic/README.md

@@ -1,51 +0,0 @@
-+++
-preset_name = "llm_request_template_anthropic"
-category = "request_template"
-mime_type = "json"
-file = "request-template-anthropic.json"
-description = "Minimal Anthropic Messages API request body"
-is_active = true
-is_public = true
-applies_to = { capability = "chat", upstream = "anthropic" }
-+++
-
-# llm / request-template-anthropic — minimal Messages API payload
-
-Canonical JSON request body for Anthropic Messages API endpoints
-(`/v1/messages`). Suitable as a test payload for seller-side
-validation and as `request_template` metadata attached to a listing.
-
-## Body
-
-```json
-{
-  "max_tokens": 100,
-  "system": "You are a helpful assistant.",
-  "messages": [
-    { "role": "user", "content": "Say hello in one sentence." }
-  ]
-}
-```
-
-## Differences from `llm_request_template`
-
-- **`max_tokens` is required** — Anthropic 4xxs without it; OpenAI
-  treats it as optional.
-- **`system` is a top-level field**, not a `role: "system"` entry in
-  `messages`. Anthropic doesn't support a system-role message.
-- No `model` field — the gateway's routing config or the listing's
-  `upstream_access_config` selects it (same convention as the OpenAI
-  template).
-
-## Conventions
-
-- `max_tokens` kept small (≤ 100) so the request completes fast.
-- Response shape is `content[0].text` rather than
-  `choices[0].message.content`. Tests should assert
-  `content[0].text` exists and is non-empty.
-
-## Versions
-
-### v1 — initial release
-
-- One user message, top-level `system`, `max_tokens = 100`, no model field.